We have data from a number of longitudinal studies that include participants’ age and a measure of cognition and carried out analyses to combine these data to examine the longitudinal decline of cognition with age. In order to do this we have used what has been described as Individual Participant Data (IPD) Meta-analysis, which has been used in published papers for a similar purpose.
We used a linear mixed model in each study separately, to examine the decline of cognition with age, with fixed effects for the intercept, age, age-squared, plus a number of control variables, and random effects for the intercept and age. (As usual, age was centred to avoid multicollinearity between the age and age-squared variables.) We then used a standard meta-analysis procedure to obtain pooled estimates of the fixed effects of the intercept, age and age-squared. These were then used to obtain a figure with a single curve displaying the variation of cognition with age based on these coefficients.
These results included in a paper submitted to a journal, and the reviewers did not question the methodology or the results, but asked that 95% confidence bands to be included in the figure. However we are not sure how to do this.
If instead of using IPD meta-analysis, we had used a single linear mixed model, with study included as either a fixed or random effect categorical variable, the SEMs or confidence intervals for the predicted mean at different ages would be given in the output of standard packages, such as SAS or in R. But we chose to use IPD meta-analysis instead, as it was the method recently used by leaders in the field, and has a number of advantages. Can anyone suggest how we might calculate confidence or SEM bands for the curve for the estimated mean based on the pooled estimates of?
Hi, Dr. Crawford.
For the IPD analysis, Ding-Geng (Din) Chen and Karl E. Peace's book Applied Meta-Analysis with R in Chapter 8 detailed the code necessary to carry out the analysis with mixed model and pooled estimates of each study included in your work.
I have a question about the needed for "traditionally" metaanalysis , because, in essence the metaanalysis is a mixed model; for the continuous variables tested in your work the best way i consider the mixed model with fixed effect for the age and other covariables (if you have them) and random effect for the patients, or use a multilevel linear model, with the patients nested in the studies.
I hope help you Dr.
If you need the book, i can send you via e-mail or onedrive.
Respectfully, Vicente.
Many thanks, Vicente. I will chase up that book which should be available from our university library. Best wishes, John.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis